DeepSARFlood: Rapid and automated SAR-based flood inundation mapping using vision transformer-based deep ensembles with uncertainty estimates

Abstract

Rapid and automated flood inundation mapping is critical for disaster management. While optical satellites provide valuable data on flood extent and impact, their real-time usage is limited by challenges such as cloud cover, limited vegetation penetration, and the inability to operate at night, making real-time flood assessments difficult. Synthetic Aperture Radar (SAR) satellites can overcome these limitations, allowing for high-resolution flood mapping. However, SAR data remains underutilized due to less availability of training data, and reliance on labor-intensive manual or semi-automated change detection methods. This study introduces a novel end-to-end methodology for generating SAR-based flood inundation maps, by training deep learning models on weak flood labels generated from concurrent optical imagery. These labels are used to train deep learning models based on Convolutional Neural Networks (CNN) and Vision Transformer (ViT) architectures, optimized through multitask learning and model soups. Additionally, we develop a novel gain algorithm to identify diverse ensemble members and estimate uncertainty through deep ensembles. Our results show that ViT-based and CNN-ViT hybrid architectures significantly outperform traditional CNN models, achieving a state-of-the-art Intersection over Union (IoU) score of 0.72 on the Sen1Floods11 test dataset, while also providing uncertainty quantification. These models have been integrated into an open-source and fully automated, Python-based tool called DeepSARFlood, and demonstrated for the Pakistan floods of 2022 and Assam (India) floods of 2020. With its high accuracy, processing speed, and ability to estimate uncertainty, DeepSARFlood is optimized for real-time deployment, processing a 1° × 1° (12,100 km²) area in under 40 s, and will complement upcoming SAR missions like NISAR and Sentinel 1-C for flood mapping.